Research On Transmission And Structure Design Of Dielectric Filter Based On Deep Learning Theory

As an important component of communication systems,the performance of microwave filter often directly affects the performance indicators of the entire communication system.Compared with the traditional metal cavity filter,the dielectric filter uses multi-stage coupling to achieve the frequency selection which has the advantages of weak dispersion and small loss.It has great prospects in the fields of mobile communications and digital satellite communications.Due to the limitation of processing technology,traditional dielectric filter design methods require repeated simulation verification to adjust the model structure.It has the disadvantages of long design period and non-reusable design processes.As the complexity of the structure increases,the design of models becomes more and more difficult.This paper proposes the use of deep neural networks with strong learning capabilities in fitting complex input-output relationships to solve this problem.Based on high-dielectric ceramic materials,the dielectric filter is designed and simulated,and the correspondence between the structure and the transmission response is explored from the simulation data,avoiding the traditional mathematical and physical calculation methods,realizing the fast prediction of the transmission response of the dielectric filter and the reverse structural design of the target transmission spectrum,which improves the calculation speed of the transmission response and the design efficiency of the model.Firstly,a band-pass dielectric filter structure working in the X-band(8-12GHz)of the communication field is proposed.It consists of periodically arranged resonance units which contains three ceramic dielectric rods arranged along the propagation direction.A microwave bandpass filter with a center frequency of 9.31 GHz and a 3dB bandwidth of 0.74GHz was obtained.Then CST software is used to build and simulate models to help a deep understanding of the resonance mechanism of the dielectric structure;Based on this,the distribution of electric field,magnetic field,and surface current were observed to analyze the influence of the structure size,material,Lattice period and the filling rate of dielectric on the filtering performance,and the characteristic responses corresponding to different structural parameters are collected as a data set for training the neural network.Then,based on deep learning theory,a fully connected neural network is designed as a forward network,which is trained using the data set obtained by the actual simulation.After 5000 iterations,the loss is reduced to 0.09,and the neural network can accurately predict the transmission spectrum of filters with unknown response on the X-band;Similarly a fully connected neural network was designed,and cascade the trained forward network as a reverse design network.After 3000 iterations,the loss value is reduced to 0.13,and the structural parameters of the filter can be accurately designed according to the target transmission response.Finally,the ideal filter curve is used to test the reverse design network.The results show that the filter designed by the reverse network is significantly different from the actual one,which is mainly because the ideal filter is practically impossible,and the data set is derived from simulation.In addition,the loss function does not consider whether the designed structure is realistic.To solve these problem,this paper introduces a weighted loss method and batch normalization method.The optimized reverse design network has greatly improved the performance on the center frequency and bandwidth,which can provide a fast and effective reference for the design of the dielectric filter and the model is realistic.